Converting protein tertiary structure into discrete tokens via vector-quantized variational autoencoders (VQ-VAEs) creates a language of 3D geometry and provides a natural interface between sequence and structure models. While pose invariance is commonly enforced, retaining chirality and directional cues without sacrificing reconstruction accuracy remains challenging. In this paper, we introduce GCPVQVAE, a geometry-complete tokenizer built around a strictly SE(3)-equivariant GCPNet encoder that preserves orientation and chirality of protein backbones. We vector-quantize rotation/translation-invariant readouts that retain chirality into a 4 096-token vocabulary, and a transformer decoder maps tokens back to backbone coordinates via a 6D rotation head trained with SE(3)-invariant objectives. Building on these properties, we train GCP-VQVAE on a corpus of 24 million monomer protein backbone structures gathered from the AlphaFold Protein Structure Database. On the CAMEO2024, CASP15, and CASP16 evaluation datasets, the model achieves backbone RMSDs of 0.4377 Å, 0.5293 Å, and 0.7567 Å, respectively, and achieves 100% codebook utilization on a held-out validation set, substantially outperforming prior VQ-VAE–based tokenizers and achieving state-of-the-art performance. Beyond these benchmarks, on a zero-shot set of 1 938 completely new experimental structures, GCP-VQVAE attains a backbone RMSD of 0.8193 Å and a TM-score of 0.9673, demonstrating robust generalization to unseen proteins. Lastly, we elaborate on the various applications of this foundation-like model, such as protein structure compression and the integration of generative protein language models. We make the GCP-VQVAE source code, zero-shot dataset, and its pretrained weights fully open for the research community: [GitHub][1]. ### Competing Interest Statement The authors have declared no competing interest. [1]: https://github.com/mahdip72/vq_encoder_decoder

Accurate identification of protein–ligand binding sites is critical for mechanistic biology and drug discovery, yet performance varies widely across ligand families and data regimes. We present a systematic prediction and evaluation framework that stratifies ligands into three settings, overrepresented (many examples), under-represented (tens of examples; few-shot), and zero-shot (unseen at training). We developed a novel three-stage, sequence-based modeling suite that progressively adds ligand conditioning and zero-shot capability, and used an evaluation frame-work to assess the suite. Stage 1 trains per-ligand predictors using a pretrained protein language model (PLM). Stage 2 introduces ligand-aware conditioning via an embedding table, enabling a single multi-ligand model. Stage 3 replaces the table with a pretrained chemical language model (CLM) operating on SMILES, enabling zero-shot generalization. We show Stage 2 improves Macro F 1 on the overrepresented test set from 0.4769 (Stage 1) to 0.5832 and outperforms sequence- and structure-based baselines. Stage 3 attains zero-shot performance ( F 1 = 0.3109) on 5 612 previously unseen ligands while remaining competitive on represented ligands. Ablations across five PLM scales and multiple CLMs reveal larger PLM backbones consistently increase Macro F 1 across all regimes, whereas scaling the CLM yields modest or inconsistent gains, which need further investigation. Our results demonstrate that zero-shot residue-level prediction from sequence and SMILES is feasible and identifies the PLM scale as the dominant lever for further advances. The code is fully open source at [GitHub][1]. ### Competing Interest Statement The authors have declared no competing interest. [1]: https://github.com/mahdip72/ProteinLigand

Accurate prediction of protein-ligand binding sites is critical for understanding molecular interactions and advancing drug discovery. Existing computational approaches often suffer from limited generality, restricting their applicability to a small subset of ligands, while data scarcity further impairs performance, particularly for underrepresented ligand types. To address these challenges, we introduce a unified model that integrates a protein language model with an autoregressive transformer for protein-ligand binding site prediction. By framing the task as a language modeling problem and incorporating task-specific tokens, our method achieves broad ligand coverage while relying solely on protein sequence input. We systematically analyze ligand-specific task token embeddings, demonstrating that they capture meaningful biochemical properties through clustering and correlation analyses. Furthermore, our multi-task learning strategy enables effective knowledge transfer across ligands, significantly improving predictions for those with limited training data. Experimental evaluations on 41 ligands highlight the model superior generalization and applicability compared to existing methods. This work establishes a scalable generative AI framework for binding site prediction, laying the foundation for future extensions incorporating structural information and richer ligand representations. The code, model, and datasets are available at https://github.com/mahdip72/prot2token. ### Competing Interest Statement The authors have declared no competing interest.

Comprehensive protein function and property prediction remains a major challenge due to the vast diversity of sequences, structural variations, and limited labeled data. Existing models are often specialized to be task-specific, requiring independent training, which limits scalability. To address this, we extend Prot2Token, a unified autoregressive framework that focuses on the post-training alignment of pre-trained protein language models (PLMs), to new applications. Our approach enables next-token prediction across new applications of protein-prediction tasks, including protein-protein structure similarity, 3D structure prediction, mutation stability, post-translational modifications (PTMs), substrate-kinase phosphorylation sites, protein-protein affinity, and protein-ion binding sites. We introduce a self-supervised pre-training stage for the decoder, enhancing model initialization and improving downstream predictions. By integrating a causal autoregressive transformer with a pre-trained ESM-2 encoder, our model effectively aligns diverse protein tasks within a single framework. Additionally, we discuss the opportunities and limitations of this approach, providing insights for future research in optimizing PLMs as a general tool for broader biological applications. Code is available on \href{https://github.com/mahdip72/prot2token}{GitHub Repository}. ### Competing Interest Statement The authors have declared no competing interest.

The diverse nature of protein prediction tasks has traditionally necessitated specialized models, hindering the development of broadly applicable and computationally efficient Protein Language Models (PLMs). In this work, we introduce Prot2Token, a unified framework that overcomes these challenges by converting a wide spectrum of protein-related predictions-from sequence-level properties and residue-specific attributes to complex inter-protein interactions-into a standardized next-token prediction format. At its core, Prot2Token employs an autoregressive decoder, conditioned on embeddings from pre-trained protein encoders and guided by learnable task tokens, to perform diverse predictions. This architecture uniquely facilitates multi-task learning, enabling general-purpose decoders to generalize across five distinct categories. We present extensive experimental validation across a variety of benchmarks, demonstrating Prot2Token's predictive power in different types of protein-prediction tasks. In 3D structure prediction, Prot2Token delivers substantial speedups (up to 1000x faster than AlphaFold2 with MSA on the same hardware) while, across other numerous tasks, matching or surpassing specialized methods. Beyond that, we introduce an auxiliary self-supervised decoder pre-training approach to improve spatially sensitive task performance. Prot2Token thus offers a step towards standardizing biological prediction into a generative interface, promising to accelerate biological discovery and the development of novel therapeutics. The code is available at https://github.com/mahdip72/prot2token .

Proteins are crucial in a wide range of biological and engineering processes. Large protein language models (PLMs) can significantly advance our understanding and engineering of proteins. However, the effectiveness of PLMs in prediction and design is largely based on the representations derived from protein sequences. Without incorporating the three-dimensional (3D) structures of proteins, PLMs would overlook crucial aspects of how proteins interact with other molecules, thereby limiting their predictive accuracy. To address this issue, we present S-PLM, a 3D structure-aware PLM, that employs multi-view contrastive learning to align protein sequences with their 3D structures in a unified latent space. Previously, we utilized a contact map-based approach to encode structural information, applying the Swin-Transformer to contact maps derived from AlphaFold-predicted protein structures. This work introduces a new approach that leverages a geometric vector perceptron (GVP) model to process 3D coordinates and obtain structural embeddings. We focus on the application of structure-aware models for protein-related tasks by utilizing efficient fine-tuning methods to achieve optimal performance without significant computational costs. Our results show that S-PLM outperforms sequence-only PLMs across all protein clustering and classification tasks, achieving performance on par with state-of-the-art methods that require both sequence and structure inputs. S-PLM and its tuning tools are available at https://github.com/duolinwang/S-PLM/ .

With the growing popularity of, and easy access to, the Internet, controlling the distribution of sensitive content such as adult images and videos is increasingly important. The abundance of images and videos available on the Internet makes it impossible to distinguish potentially harmful content (adult images or videos) from benign content (non-adult images or videos) through manual monitoring. To address this challenge, This paper presents a hybrid architecture combining Convolutional Neural Networks (CNN) and Transformers to address the challenge of recognizing adult images and videos on the Internet. Due to the vast amount of such content, manual monitoring is impractical. With the growing popularity of, and easy access to, the Internet, controlling the distribution of sensitive content such as adult images and videos is increasingly important. The abundance of images and videos available on the Internet makes it impossible to distinguish potentially harmful content (adult images or videos) from benign content (non-adult images or videos) through manual monitoring. To address this challenge, the proposed method integrates transformer blocks after a convolutional encoder to enhance the representation learning process, focusing on essential features. By examining and selecting the optimal trade-off between the number of layers in the transformer blocks, the proposed method selects essential features rather than minor ones. We show that our model is also effective as a powerful feature extractor for video frames in video classification. Experiments demonstrate that our method outperforms state-of-the-art methods in adult content video classification on the NPDI, Pornography-800, and Pornography-2000 datasets, with minimal computational overhead. Adding transformer blocks on top of the CNN backbone introduces only a few additional parameters to the whole architecture compared to the convolutional backbone, resulting in minimal extra computational cost. Statistical analyses, including non-parametric tests, validate the model’s performance, confirming its robustness.

Protein targeting, often guided by targeting peptides, is a critical biological process that directs proteins to their specific cellular destinations, ensuring proper cellular functionality and organization. Accurate classification and detection of targeting peptides are fundamental to understanding protein sorting mechanisms. This study introduces MULoc-Target, a novel deep-learning method designed to detect and classify targeting peptides in eukaryotic proteins. To support its development and evaluation, we curated a benchmark dataset comprising eight types of eukaryotic targeting peptides with manually curated annotations. Comprehensive evaluations on this dataset and external datasets from the literature demonstrate that MULoc-Target achieves state-of-the-art or competitive performance in detecting and classifying targeting peptides. Additionally, it enables the extraction of enriched motif patterns, offering valuable insights into their properties and the underlying targeting mechanisms. The identified motifs align closely with established biological features, further validating MULoc-Target's capabilities. A web server for MULoc-Target is integrated into our MULocDeep localization suite as a new toolkit, publicly accessible at https://mu-loc.org/MULoc-Target, and the inference code is available at https://github.com/yuexujiang/MULoc-Target.

This paper proposes a versatile tokenization method and introduces Prot2Token, a model that combines autoregressive language modeling with protein language models (PLMs) to tackle various protein prediction tasks using protein sequences. Leveraging our tokenization method, Prot2Token adapts existing PLMs for multiple tasks such as protein-level prediction, residue-level prediction, and protein-protein interaction prediction through next-token prediction of tokenized target label sequences. By incorporating prompt tokens into the decoder, Prot2Token enables multi-task training in a single end-to-end session. Our results demonstrate that Prot2Token not only matches the performance of specialized models across various tasks but also paves the way for integrating protein tasks with large language models (LLMs), representing an important step towards creating general-purpose PLMs for advanced protein language processing (PLP). Additionally, we use Prot2Token to develop S-ESM, a structure-aware version of the ESM model, which achieves competitive performance with state-of-the-art methods in 3D structure-related tasks using only protein sequences. Code is available at: . ### Competing Interest Statement The authors have declared no competing interest.

Proteins play an essential role in various biological and engineering processes. Large protein language models (PLMs) present excellent potential to reshape protein research by accelerating the determination of protein function and the design of proteins with the desired functions. The prediction and design capacity of PLMs relies on the representation gained from the protein sequences. However, the lack of crucial 3D structure information in most PLMs restricts the prediction capacity of PLMs in various applications, especially those heavily dependent on 3D structures. To address this issue, we introduce S-PLM, a 3D structure-aware PLM that utilizes multi-view contrastive learning to align the sequence and 3D structure of a protein in a coordinated latent space. S-PLM applies Swin-Transformer on AlphaFold-predicted protein structures to embed the structural information and fuses it into sequence-based embedding from ESM2. Additionally, we provide a library of lightweight tuning tools to adapt S-PLM for diverse protein property prediction tasks. Our results demonstrate S-PLM's superior performance over sequence-only PLMs on all protein clustering and classification tasks, achieving competitiveness comparable to state-of-the-art methods requiring both sequence and structure inputs. S-PLM and its lightweight tuning tools are available at https://github.com/duolinwang/S-PLM/ .

In this paper, at first, the impact of ImageNet pre-training on fine-grained Facial Emotion Recognition (FER) is investigated which shows that when enough augmentations on images are applied, training from scratch provides better result than fine-tuning on ImageNet pre-training. Next, we propose a method to improve fine-grained and in-the-wild FER, called Hybrid Multi-Task Learning (HMTL). HMTL uses Self-Supervised Learning (SSL) as an auxiliary task during classical Supervised Learning (SL) in the form of Multi-Task Learning (MTL). Leveraging SSL during training can gain additional information from images for the primary fine-grained SL task. We investigate how proposed HMTL can be used in the FER domain by designing two customized version of common pre-text task techniques, puzzling and in-painting. We achieve state-of-the-art results on the AffectNet benchmark via two types of HMTL, without utilizing pre-training on additional data. Experimental results on the common SSL pre-training and proposed HMTL demonstrate the difference and superiority of our work. However, HMTL is not only limited to FER domain. Experiments on two types of fine-grained facial tasks, i.e., head pose estimation and gender recognition, reveals the potential of using HMTL to improve fine-grained facial representation.

Post-translational modifications (PTMs) have key roles in extending the functional diversity of proteins and, as a result, regulating diverse cellular processes in prokaryotic and eukaryotic organisms. Phosphorylation modification is a vital PTM that occurs in most proteins and plays a significant role in many biological processes. Disorders in the phosphorylation process lead to multiple diseases, including neurological disorders and cancers. The purpose of this review is to organize this body of knowledge associated with phosphorylation site (p-site) prediction to facilitate future research in this field. At first, we comprehensively review all related databases and introduce all steps regarding dataset creation, data preprocessing, and method evaluation in p-site prediction. Next, we investigate p-site prediction methods, which are divided into two computational groups: algorithmic and machine learning (ML). Additionally, it is shown that there are basically two main approaches for p-site prediction by ML: conventional and end-to-end deep learning methods, both of which are given an overview. Moreover, this review introduces the most important feature extraction techniques, which have mostly been used in p-site prediction. Finally, we create three test sets from new proteins related to the released version of the database of protein post-translational modifications (dbPTM) in 2022 based on general and human species. Evaluating online p-site prediction tools on newly added proteins introduced in the dbPTM 2022 release, distinct from those in the dbPTM 2019 release, reveals their limitations. In other words, the actual performance of these online p-site prediction tools on unseen proteins is notably lower than the results reported in their respective research papers.

In recent years, the main problem in e-learning has shifted to personalization of learning environment by Intelligent Tutoring Systems (ITSs). Therefore, by designing personalized teaching models, learners are able to have a successful and satisfying experience in achieving their learning goals. Affective Tutoring Systems (ATSs) are some kinds of ITS that can recognize and respond to affective states of learners. In this study, we have designed, implemented, and evaluated an ATS named ATTENDEE (AffecTive Tutoring system based on facial EmotioN recognition and heaD posE Estimation) to personalize the learning environment based on the facial emotions recognition, head pose estimation, and cognitive style of learners. First, a unit called Intelligent Analyzer (IA) created which was responsible for recognizing facial expression and head angles of learners. Next, the ATS was built which mainly made of two units: ITS, IA. Results indicated that with the ATS, participants needed less efforts to pass the tests. In other words, we observed when the IA unit was activated, learners could pass the final tests in fewer attempts than those for whom the IA unit was deactivated. In addition, we have examined the effect of the IA unit on the educational achievement and satisfaction of learners.

Protein ubiquitination is a critical post-translational modification (PTMs) involved in numerous cellular processes. Identifying ubiquitination sites (Ubi-sites) on proteins offers valuable insights into their function and regulatory mechanisms. Due to the cost- and time-consuming nature of traditional approaches for Ubi-site detection, there has been a growing interest in leveraging artificial intelligence for computer-aided Ubi-site prediction. In this study, we collected experimentally verified Ubi-sites of human proteins from the dbPTM database, then conducted comprehensive state-of-the art computational methods along with standard evaluation metrics and a proper validation strategy for Ubi-site prediction. We presented the effectiveness of our framework by comparing ten machine learning (ML) based approaches in three different categories: feature-based conventional ML methods, end-to-end sequence-based deep learning (DL) techniques, and hybrid feature-based DL models. Our results revealed that DL approaches outperformed the classical ML methods, achieving a 0.902 F1-score, 0.8198 accuracy, 0.8786 precision, and 0.9147 recall as the best performance for a DL model using both raw amino acid sequences and hand-crafted features. Interestingly, our experimental results disclosed that the performance of DL methods had a positive correlation with the length of amino acid fragments, suggesting that utilizing the entire sequence can lead to more accurate predictions in future research endeavors. Additionally, we developed a meticulously curated benchmark for Ubi-site prediction in human proteins. This benchmark serves as a valuable resource for future studies, enabling fair and accurate comparisons between different methods. Overall, our work highlights the potential of ML, particularly DL techniques, in predicting Ubi-sites and furthering our knowledge of protein regulation through ubiquitination in cells.

Post-translational modifications (PTMs) have vital roles in extending the functional diversity of proteins and as a result, regulating diverse cellular processes in prokaryotic and eukaryotic organisms. Phosphorylation modification is a vital PTM that occurs in most proteins and plays significant roles in many biological processes. Disorders in the phosphorylation process lead to multiple diseases including neurological disorders and cancers. At first, this study comprehensively reviewed all databases related to phosphorylation sites (p-sites). Secondly, we introduced all steps regarding dataset creation, data preprocessing and method evaluation in p-sites prediction. Next, we investigated p-sites prediction methods which fall into two computational and Machine Learning (ML) groups. Additionally, it was shown that there are basically two main approaches for p-sites prediction by ML: conventional and End-to-End learning, which were given an overview for both of them. Moreover, this study introduced the most important feature extraction techniques which have mostly been used in ML approaches. Finally, we created three test sets from new proteins related to the 2022th released version of the dbPTM database based on general and human species. After evaluating available online tools on the test sets, results showed that the performance of online tools for p-sites prediction are quite weak on new reported phospho-proteins.

The cost of head pose labeling is the main challenge of improving the fine-grained Head Pose Estimation (HPE). Although Self-Supervised Learning (SSL) can be a solution to the lack of huge amounts of labeled data, its efficacy for fine-grained HPE is not yet fully explored. This study aims to assess the usage of SSL in fine-grained HPE based on two scenarios : (1) using SSL for weights pre-training procedure , and (2) leveraging auxiliary SSL losses besides HPE. We design a Hybrid Multi-Task Learning (HMTL) architecture based on the ResNet50 backbone in which both strategies are applied. Our experimental results reveal that the combination of both scenarios is the best for HPE. Together, the average error rate is reduced up to 23.1% for AFLW2000 and 14.2% for BIWI benchmark compared to the baseline. Moreover, it is found that some SSL methods are more suitable for transfer learning, while others may be effective when they are considered as auxiliary tasks incorporated into supervised learning. Finally, it is shown that by using the proposed HMTL architecture, the average error is reduced with different types of initial weights: random, ImageNet and SSL pre-trained weights.

In recent years, the main problem in e-learning has shifted from analyzing content to personalization of learning environment by Intelligence Tutoring Systems (ITSs). Therefore, by designing personalized teaching models, learners are able to have a successful and satisfying experience in achieving their learning goals. Affective Tutoring Systems (ATSs) are some kinds of ITS that can recognize and respond to affective states of learner. In this study, we designed, implemented, and evaluated a system to personalize the learning environment based on the facial emotions recognition, head pose estimation, and cognitive style of learners. First, a unit called Intelligent Analyzer (AI) created which was responsible for recognizing facial expression and head angles of learners. Next, the ATS was built which mainly made of two units: ITS, IA. Results indicated that with the ATS, participants needed less efforts to pass the tests. In other words, we observed when the IA unit was activated, learners could pass the final tests in fewer attempts than those for whom the IA unit was deactivated. Additionally, they showed an improvement in terms of the mean passing score and academic satisfaction.

In this paper, at first, the impact of ImageNet pre-training on Facial Expression Recognition (FER) was tested under different augmentation levels. It could be seen from the results that training from scratch could reach better performance compared to ImageNet fine-tuning at stronger augmentation levels. After that, a framework was proposed for standard Supervised Learning (SL), called Hybrid Learning (HL) which used Self-Supervised co-training with SL in Multi-Task Learning (MTL) manner. Leveraging Self-Supervised Learning (SSL) could gain additional information from input data like spatial information from faces which helped the main SL task. It is been investigated how this method could be used for FER problems with self-supervised pre-tasks such as Jigsaw puzzling and in-painting. The supervised head (SH) was helped by these two methods to lower the error rate under different augmentations and low data regime in the same training settings. The state-of-the-art was reached on AffectNet via two completely different HL methods, without utilizing additional datasets. Moreover, HL's effect was shown on two different facial-related problem, head poses estimation and gender recognition, which concluded to reduce in error rate by up to 9% and 1% respectively. Also, we saw that the HL methods prevented the model from reaching overfitting.